Strongly Guided Pixel-level Supervised Contrast for Polyp Segmentation

The current polyp segmentation methods mainly use the saliency map to obtain the uncertain region, foreground region, and background region of the polyp image, and then they learn the semantic information from each other, to enhance the edge segmentation ability of the network. However, there is great instability in the quality of the saliency map and the error information brought by low-quality saliency maps will interfere with the segmentation ability of the network. To this end, this paper proposes a strongly guided pixel-level supervised contrastive learning method (SGP-SCL), which enhance the model to identify the polyp boundary by strengthening foreground and background guidance for polyp boundary. Specifically, the SGPS-CL method fully utilizes the ground truth label to obtain high-confidence and representative samples to guide the learning of boundary regions with low confidence, thus reducing the impact of the instability of the preliminary prediction probability map quality on the network performance. Experiments are conducted on CVC-300, CVCClinicDB, Kvasir, CVC-ColonDB, and ETIS polyp segmentation datasets, and the proposed method achieves competitive results.